AgentPool defines 29 tools with consistent verb-noun naming (get_, spawn_, observe_, etc.) and detailed descriptions (avg ~150 chars). All tools have input schemas with typed parameters and descriptions. However, output schemas are not explicitly documented in the source, and several tools lack clarity on return structure. Error handling is minimal, most tools delegate to manager methods without recovery guidance. Security considerations (lockdown flags, lease management) are present but not systematized. Composition is strong: tools chain well (spawn → observe → collect), and parameter naming is consistent. The server follows good patterns for stateless operation and tool annotations but falls short of A-grade polish in documentation completeness.

• Document output schemas for all tools. For each tool, add a 'Returns' section specifying field names, types, and structure. Example: 'Returns: {session_id: string, state: string, created_at: ISO8601, artifacts: {result: string, transcript: string, diff: string}}'.
• Add error classification to tool descriptions. State which errors are retryable (e.g., 'timeout_seconds exceeded'), which are user-fixable (e.g., 'invalid session_id'), and which are fatal. Include recovery guidance: 'If session not found, call list_sessions() to verify the ID.'
• Implement confirmation pattern for destructive tools. Add a 'dry_run' parameter to terminate_worker and cleanup_worktree, or create separate 'preview_terminate_worker' tools that return what would be deleted without executing.
• Formalize parameter constraints as JSON Schema enums and numeric bounds. Convert detail parameter to enum: {type: 'string', enum: ['summary', 'excerpt', 'full']}. Add minValue/maxValue to timeout_seconds, max_lines, limit, offset.
• Standardize pagination across list tools. All list_* tools should accept limit (1 - 100, default 20) and offset (default 0), and return {items: [...], total: number, has_more: boolean}. Add pagination to list_file_leases and list_worktrees.